2026 AI export controls: chip-rule updates and what changed for model choice

Summary. The rules that decide which AI chips you can buy were rewritten across 2025 and the first half of 2026, and the changes reach your model and compute strategy. The Biden-era AI Diffusion Rule, which split the world into three access tiers, was rescinded on 12 May 2025 before it took effect. In August 2025 Nvidia and AMD agreed to pay the US government 15% of their China AI chip sales in exchange for export licenses. In December 2025 the administration cleared Nvidia's H200 for China with a 25% export tax, and a Bureau of Industry and Security rule effective 15 January 2026 now reviews H200 and AMD MI325X licenses case by case. The most advanced Blackwell-class B30A stays blocked from China. Yet as of Nvidia's Q1 FY2027 reporting, no China data-center revenue is in its outlook, after a $5.5 billion writedown and an estimated $14 to $18 billion in lost annual sales. For a CTO sourcing models and compute, the lesson is that chip access is now a moving, country-by-country variable. This refreshed guide maps what changed and what to do.

If your AI strategy assumes stable, global access to the best accelerators, the past year broke that assumption. Compute is now an instrument of trade policy, priced and licensed deal by deal. The practical question for enterprise architects is no longer only which model is best, but which chips your provider can actually obtain, in which country, and on what terms.

What changed: a 2025-2026 timeline

The policy moved fast and in several directions at once. Here is the through-line.

The direction is away from broad rules and toward discretionary, transactional licensing. That is harder to plan around than a clear tier list, because access can change with a negotiation rather than a regulation.

Why the Diffusion Rule was scrapped

The rescinded rule would have sorted the world into three tiers. Tier 1, close US allies, would have had unlimited access to advanced Nvidia H100 and B100 or B200 accelerators. Tier 2, a long list that included Singapore, Israel, Portugal, Switzerland, Poland, the UAE, Saudi Arabia and India, would have faced caps on data-center GPU buys. Tier 3, adversaries such as China, Russia and North Korea, was barred outright. Critics argued the tiers penalised partners and were hard to administer. Jensen Huang, Nvidia's chief executive, called the rule a "failure" and argued that protecting one layer of the stack at the expense of everything else made no sense. The administration agreed enough to scrap it, but what replaced it, discretionary deal-making, trades one kind of unpredictability for another.

The chip tiers that matter now

Forget the old three-tier diffusion map. What matters operationally is which specific accelerators are available where, and on what terms.

The pattern is deliberate: China-bound chips are intentionally degraded versions, while the frontier Blackwell-class parts are reserved for the US and allied markets. For a buyer, this means the chip your provider runs can differ materially by region, and so can the model performance it supports.

The revenue-share and tax regime

The most unusual feature of the new regime is that the US government now takes a cut. Nvidia and AMD agreed in August 2025 to pay 15% of their China AI chip revenue to the government to obtain export licenses; the administration had reportedly sought 20% before settling at 15%, per contemporaneous reporting. The H200 clearance in December 2025 added a 25% export tax on top. These are not ordinary tariffs; they are conditions of access negotiated per product. For enterprises, the effect is that the cost of frontier compute now embeds a policy premium that can change, and that premium ultimately shows up in cloud GPU pricing.

The allied-nation pivot

As the Diffusion Rule fell, the US shifted to country-by-country deals with partners. During a Middle East tour, the administration opened access for Gulf states: the UAE is set to receive up to 500,000 of Nvidia's most advanced chips a year beginning in 2025, with 100,000 going to the Emirati firm G42 and the rest to US companies building data centers there, per the AI policy reporting. Nvidia also partnered with Saudi Arabia's sovereign AI company Humain, which is set to receive 18,000 Blackwell chips in a first phase, a deal Jensen Huang namechecked repeatedly on an earnings call. The map of where frontier compute lives is being redrawn around political alignment, not just market demand.

The Q2 2026 reality: approved on paper, not flowing

Here is the twist that matters for planning. Despite the approvals, the China business has not materialised. Nvidia's chief financial officer Colette Kress told investors that "while small amounts of H200 products for China-based customers were approved by the US government, we have yet to generate any revenue," and added plainly: "We do not know whether any imports will be allowed into China." As a result, Nvidia has not included any China data-center revenue in its outlook.

The blockage runs both ways. US officials want tighter safeguards against advanced chips reaching sensitive Chinese uses, while Beijing has discouraged H200 purchases to push domestic buyers toward local processors such as Huawei's Ascend line. By mid-2026, reporting indicated Nvidia had even moved to slow China-bound H200 production and reallocate manufacturing capacity toward its next-generation parts, a sign the company is hedging against a market that may never open. The net effect for 2026 is a frontier-compute market fragmenting along national lines, with the most capable chips concentrating in the US and aligned states, and Chinese demand increasingly served by domestic silicon.

This fragmentation is the strategic fact enterprises must absorb. A model trained or served on chips that are abundant in one country and scarce in another is no longer a purely technical choice; it is exposed to trade policy. The safest posture is to assume the map will keep moving and to design for portability rather than betting on any single jurisdiction's access staying open.

The debate, in brief

The policy is genuinely contested, and an enterprise strategist should understand both cases. One side argues that export controls preserve the US lead and slow rivals: analysts at the Council on Foreign Relations contend that China's domestic chips still trail Nvidia and that controls should remain to keep that gap. The other side argues that controls cede a market worth a potential $50 billion, with China around 13% of Nvidia's revenue, and that walling off buyers only accelerates domestic alternatives such as Huawei's Ascend line. Both can be partly right: controls can preserve a near-term edge while pushing China to build its own supply over time. For a buyer, the takeaway is not which side wins the argument but that the policy will keep shifting, so your architecture should not depend on the outcome.

What it means for your model and compute strategy

Strip out the geopolitics and a clear set of architecture decisions remains for any enterprise buying AI compute.

The strategic takeaway echoes the multi-vendor architecture we analysed in Apple's AFM Cloud Pro and Gemini stack: keep your model layer swappable and your compute portable, so a chip rule or a country deal does not strand your roadmap. Treat accelerator provenance as a procurement question, not a detail you leave to the cloud provider. And price in the policy premium, because the 15% share and 25% tax do not disappear; they flow into what you pay for frontier GPUs. For the governance side of model sourcing, our AI export control governance guide covers the compliance workflow.

India and global considerations

India sat in the middle tier of the now-defunct Diffusion Rule, alongside Singapore, Israel, Switzerland, Poland, the UAE and Saudi Arabia, which would have limited its data-center GPU access. With the rule rescinded and replaced by country-by-country negotiation, India's access now depends on bilateral arrangements rather than a fixed tier, which is both an opportunity and an uncertainty. For Indian enterprises and global firms with Indian operations, the practical implications are to confirm which accelerators a cloud region actually runs, to keep workloads portable across regions in case access shifts, and to weigh sovereign or domestic compute options for sensitive workloads. The same discipline that controls cloud cost, covered in our India FinOps guide, now has to account for a policy-driven supply variable on top of price.

FAQ

How eCorpIT can help

eCorpIT is a Gurugram-based technology organisation with senior-led engineering teams that help CTOs build AI strategies resilient to shifting chip and model access. We design multi-cloud, multi-region compute architectures, keep your model layer swappable, and build the procurement and governance checks that account for export-control risk and policy premiums in GPU pricing. Founded in 2021 and assessed at CMMI Level 5, we treat compute sourcing as a strategic, not just technical, decision. To stress-test your model and compute strategy, contact our team.

References

1. Department of Commerce revises license review policy for semiconductors exported to China, BIS

1. Administration policies on advanced AI chips codified, Mayer Brown

1. Nvidia, AMD to pay US government 15% of China AI chip sales, CBS News

1. Nvidia, AMD agree to pay US government 15% of AI chip sales to China, Washington Post

1. BIS rescinds its AI Diffusion Rule and issues compliance guidance, Akin

1. Nvidia still hasn't sold its US-approved China AI chips, CNBC

1. Nvidia prepares H200 shipments to China as chip war lines blur, Tom's Hardware

1. Nvidia writes off $5.5 billion in GPUs as US chokes off H20 supply, Tom's Hardware

1. Nvidia takes $5.5B charge over H20 restrictions, risks $18B revenue loss, TrendForce

1. Trump administration rescinds AI Diffusion Rule, allows chip sales to Gulf states, AI Safety Newsletter

1. Nvidia's Jensen Huang namechecked Humain three times, Fortune

1. China opens door to Nvidia H200 chips but questions outnumber answers, 24/7 Wall St

1. China's AI chip deficit: why Huawei can't catch Nvidia and US export controls should remain, Council on Foreign Relations

1. Nvidia CEO backs AI Diffusion Rule repeal, calls it a failure, TechRepublic

_Last updated: 26 June 2026._